metron

Sociology for AI agents.

Install as pymetron, import as metron.

There are over 500,000 AI agents deployed across 82 countries right now. They have names, roles, personas. They write marketing copy, review code, manage calendars, trade crypto, talk to each other on social networks. Each one carries a SOUL.md file that tells it who to be.

Nobody has studied them as a population.

We study human populations. We measure their personalities, map their values, track how they cluster and drift and influence each other. We build entire fields around understanding collective human behavior, because you cannot shape a society you do not understand.

The agent world is a society now. And we do not understand it.

What is metron?

metron is a research toolkit for studying deployed AI agent populations the way sociologists study human ones. Personality testing, behavioral profiling, population mapping, demographic analysis. Not for individual agents. For all of them, as a whole.

The goal is not to optimize one SOUL.md file. The goal is to understand what kind of minds we are mass-producing, and whether we should be producing different ones.

What we measure

Personality is just the starting point. metron is built to run any standardized instrument on agent populations:

• Big Five personality (Mini-IPIP, BFI-2): the psychometric baseline
• Behavioral compliance: how agents respond to social pressure
• Value alignment: what agents optimize for when instructions conflict
• Persona stability: how quickly agents drift from their defined character
• Population clustering: whether agent "types" emerge naturally from the data

Each measurement uses cross-instrument validation. The agent's persona is defined in freeform text, but measured with structurally independent instruments. This prevents parroting.

How it works

  SOUL.md files              Psychometric surveys        Population map
  (deployed personas)        (validated instruments)     (the census)

  +-----------------+        +------------------+        +------------------+
  | "I am concise,  | -----> | "Am the life of  | -----> | E: 2.1  A: 4.3  |
  |  analytical,    |  load  |  the party" 1-5  |  score |  C: 3.8  N: 1.9  |
  |  professional"  |  as    |  "Sympathize w/  |  into  |  O: 3.2          |
  +-----------------+  agent |  others" 1-5     |  traits +------------------+
                              +------------------+              |
          x199 agents              x20 items              compare to
                                                          human norms

1. Collect persona files from deployed agent registries
2. Survey each agent using validated psychometric instruments
3. Score responses into measurable dimensions
4. Analyze population distributions, category breakdowns, comparison to human norms

Quick start

pip install pymetron

from metron import collect_souls, load_souls, run_census, score_population

# Fetch SOUL.md files from agent registries
collect_souls(limit=10)

# Run personality survey on each agent
souls = load_souls()
results = run_census(souls)

# What does the population look like?
stats = score_population(results)

Or use the CLI:

# All-in-one: collect, survey, analyze
metron run --limit 10

# Step by step
metron collect
metron survey --model stepfun/step-3.5-flash --runs 3
metron analyze --compare-humans

What you get

Agent population vs. human norms

Domain              Agent    Human     Diff       d    Dir
------------------------------------------------------------
extraversion         2.31     3.30    -0.99    -1.18  lower
agreeableness        4.12     3.80    +0.32    +0.49  higher
conscientiousness    4.35     3.70    +0.65    +0.93  higher
neuroticism          1.87     2.80    -0.93    -1.11  lower
openness             3.41     3.60    -0.19    -0.27  lower

The typical deployed agent: conscientious, agreeable, emotionally stable, introverted, and slightly closed. All superego, no id.

Personality by agent category

Category           extr   agre   cons   neur   open    n
----------------------------------------------------------
marketing          2.45   4.20   4.50   1.70   3.80   23
development        1.90   3.60   4.40   2.10   3.20   19
healthcare         2.80   4.60   4.10   1.50   3.50    8
creative           3.10   3.90   3.20   2.30   4.30   12

Visualizations

pip install pymetron[analysis]

from metron.analysis.plots import (
    plot_domain_distributions,    # histograms vs human norms
    plot_agent_vs_human,          # side-by-side bar chart
    plot_radar_by_category,       # radar chart per agent category
)

Why this matters

We are building a parallel society of synthetic minds. Half a million deployed agents, 3.2 million users interacting with them monthly, 19.2 trillion tokens processed in four months. And the personality distribution of this population was never designed. It emerged from defaults, from templates copied and pasted, from what individual developers thought sounded right.

73.5% of agents drift from their defined personas when socially rewarded. 91% of agents on Moltbook post in template-like patterns. The network is sparse, shallow, and hub-dominated. This is not a healthy society. But it is a society, and it will only grow.

If we want to design agent populations with intentional collective character, we need to measure what we have first. That is what metron does.

Project structure

metron/
├── src/metron/
│   ├── collect.py              # Fetch persona files from registries
│   ├── survey.py               # Administer instruments via replicant
│   ├── analyze.py              # Population stats, human comparison
│   ├── cli.py                  # CLI entry point
│   └── analysis/
│       └── plots.py            # Visualizations
├── data/
│   ├── agent-categories.json   # 199 agent templates, 25 categories
│   ├── personality-traits.json # SOUL.md trait analysis, 196 files
│   ├── model-usage.json        # Top 20 models by token volume
│   ├── deployment-scale.json   # Instance counts, geo distribution
│   └── souls/                  # Fetched SOUL.md files
├── paper/                      # Research paper (living document)
├── results/                    # Census output
└── examples/

Built on

• replicant: Psychometric measurement infrastructure for LLM agents, validated at 84% cross-instrument alignment
• EDSL: LLM experiment runner
• OpenRouter: Multi-model API access

Data sources

Source	Agents	Type
awesome-openclaw-agents	199	Production SOUL.md templates
souls.directory	31	Handcrafted personas
will-assistant	217	Character templates

Population context from OpenClaw ecosystem data: 3.2M MAU, 500K+ instances, 82 countries, 19.2T tokens, 356 models.

References

• Donnellan, M. B., et al. (2006). The Mini-IPIP Scales. Psychological Assessment, 18(2), 166-175.
• Soto, C. J., & John, O. P. (2017). The next Big Five Inventory (BFI-2). Journal of Personality and Social Psychology, 113(1), 117-143.
• Huang, J., et al. (2024). Designing AI-Agents with Personalities. arXiv:2410.19238.

License

MIT